Research On Image Recognition Based On Cost-sensitive Multi-granularity Three-way Decision

The cost-sensitive multi-granularity three-way decision model is based on costsensitive learning and granular computing.it can effectively solve complex problems in various fields and make great breakthroughs in data mining,medical detection,intelligent control,pattern recognition,human-computer interaction and other aspects.Different from classical decision models,cost-sensitive multi-granularity three-way decision can not only stratify,classify and cluster data,but also consider the cost loss of granularity space while rapidly processing large-scale data,thus improving the accuracy of decision.The combination of autoencoder network and cost-sensitive multi-granularity three-way decision can effectively avoid the shortcomings of the network itself,but using the traditional autoencoder network cost-sensitive multi-granularity three-way decision model in image recognition will increase the cost and reduce the performance of the model.Therefore,the main research work of this thesis is as follows:Aiming at the problems that the existing optimal granularity selection method of autoencoder network leads to poor network feature extraction ability,high error recognition cost and high testing cost,a granule layer selection method based on mini-batch gradient descent is proposed.This method reconstructs the multi-granularity space of the three-way decision through changing the granularity selection strategy,trains the model parameters by dividing the samples into finer granularity scales,and mines the important characteristic information,thereby promoting the construction of multi-granularity space towards the direction of reaching the finest granularity space faster.The feature information extracted from each granularity layer is sent to the cost-sensitive classifier to obtain the posterior probability,then the cost-sensitive three-way decision is made based on the lowest cost to identify the sample category.The optimal granularity selection is the key to constructing multi-granularity features in the network,the experimental results show that the method proposed in this thesis can effectively improve the decision accuracy of the cost-sensitive multi-granularity three-way decision model.With respect to the problem of the multi-granularity feature definition of the nonlinear feature extraction network does not satisfy the total order relationship of the three-way decision,resulting in redundant decision steps in the multi-granularity three-way decision process,which increases the decision cost and reduces the performance of the decision model,a deep belief network cost-sensitive multi-granularity three-way decision model is proposed.Based on the mini-batch gradient descent,the search domain is proposed to represents the retrieved feasible domain,analyzes the internal correlation between the reconstruction error and the search domain,and proves that the search domain set corresponding to the monotonically decreasing reconstruction error set satisfies the total order relationship,finally,redefine the multi-granularity feature set and construct a multigranularity space that satisfies the total order relationship for three-way decision.Under the same experimental environment,the deep belief network cost-sensitive multi-granularity three-way decision model reduces the decision cost on different data sets,and improves the performance and accuracy of the model.